An ecommerce agency can sell a bounded AI product recommendation audit built around evidence the client can inspect.
Start with one client product, one buyer question, and one answer engine.
Answer in brief
A client-ready recommendation audit has four stages:
- Page evidence: record the public product facts the buyer question depends on.
- Answer evidence: record what a named engine returns for that exact question.
- One correction: fix the single evidence gap that matters most, with the client’s permission.
- Rerun: run the unchanged question again and report the answer delta.
The client needs the evidence behind each stage. A readiness score, a screenshot, or a claim that a product is “AI optimized” cannot show the work.
Which client product should an agency test first?
Choose a product with a clear buyer job and enough commercial value to justify the work. Skip branded prompts. They only test whether an engine can repeat a name the user already supplied.
Good prompts name a purchase constraint:
What braiser works on induction and can go in a 500 degree oven?
What loose-leaf black tea should I buy for a strong everyday breakfast brew?
What spark plug wire set fits a classic air-cooled engine and keeps the original look?
Each of those ties to a specific product, a set of factual requirements, and a comparison set.
What should the agency establish before changing the page?
Record two separate baselines before anything on the store changes.
Record the product page first: identity, price, availability, identifiers, structured data, policies, and the product claims the question depends on.
Then record the answer: engine, exact prompt, date, mentioned products, merchant citations, and a hash of the response. The client should be able to see what the engine returned without anyone treating it as stable placement.
Our August 13, 2026 test covered 15 selected public Shopify products in logged-out Perplexity sessions.
- MEASURED: thirteen product pages passed every deterministic product check.
- OBSERVED: 13 target products were absent from their answer, and the rendered answers cited none of the target merchant domains.
- INFERRED: some clients need a factual correction on the page. Others need comparison, authority, or buyer-job evidence. Those observations separate the two jobs without establishing causation.
The field review found one specific page change to evaluate. No merchant change or unchanged-prompt rerun followed that run.
How does the agency choose what to improve first?
Use the smallest correction that addresses the buyer question.
If the product’s facts are missing or inconsistent, correct the factual surface first. If the page is already technically complete, look at whether it explains the buyer job and separates itself truthfully from the alternatives the engine named.
Both cases showed up in the cohort. An anonymized cookware product had a concrete deterministic gap and was absent from an answer that named four alternatives. An anonymized finishing-oil product passed every deterministic check and was still absent from its buyer question. The first is a page fix. The second needs evidence about the buyer job and the comparison, which is slower work and harder to attribute.
Do not change schema, product copy, site architecture, or off-site distribution before the same rerun. The client will not know which one mattered.
What should the client report contain?
- store, product URL, and target product
- exact buyer question and named engine
- verified, inferred, and runtime-required product evidence
- baseline answer hash, mentions, citations, and compared products
- one approved correction and a record of the surface that changed
- unchanged-prompt rerun evidence
- gained, unchanged, or lost mention and citation status
Keep commercial outcomes out of it. A changed answer does not establish qualified traffic, conversion, or revenue. Those need attribution well beyond this audit.
Can an agency sell this before it has a before-and-after case?
Yes, as a bounded pilot with accurate expectations. The first goal is one permissioned correction and one unchanged-prompt rerun. Until that exists, the agency has a measurement method and baseline evidence, not proof that its work changes placement.
The cohort records recommendation misses in one named-engine test. It says nothing yet about willingness to pay or a repeatable relationship between a page fix and placement.
FAQ
How many products should be in the first client audit?
One to three. Choose distinct buyer jobs and keep each prompt fixed.
Should an agency combine ChatGPT, Gemini, and Perplexity into one score?
No. Record each engine separately. They can use different sources and return different products for the same question.
Does a client need to install an app before the audit?
No. The Recommendation Audit reads one public Shopify product page. Any later store change needs the client’s authority and happens outside Colter.
What is the first successful agency outcome?
A permissioned correction followed by an unchanged-prompt rerun with a clearly recorded answer delta. A reply, a meeting, or a higher score is not product proof.
Run a client Recommendation Audit
Review the Recommendation Audit proof method and evidence documentation before presenting the result to a client.
Evidence record
- Date: August 13, 2026
- Engine: Perplexity Search, logged out, public default experience
- Cohort: 15 selected public Shopify products
- Method: one prewritten unbranded category prompt per product; one observation per prompt
- Observed: 13 targets omitted; two mentioned; zero target merchant-domain citations
- Product state: 13 pages passed deterministic product checks; 11 of those were absent
- Fix boundary: one actionable page-fix candidate identified; no merchant change or unchanged-prompt rerun performed
- Limitation: no permissioned merchant fix-to-rerun result, stable-placement or causation evidence, cross-engine rate, traffic, conversion, or revenue evidence
The underlying answer records are retained internally. Public examples are anonymized to avoid identifying the merchants or products.